redaction

A follow-up post summarizing the 4 hardest edge cases that break AI agent data sanitization in production — schema drift, allowlist rot, answer-quality degradation, and cross-tool composite PII — plus the hybrid architecture lessons learned from each.

The tweet discusses how an Ant employee agrees with a safety report but is too busy to share details, highlighting understaffing in AI safety, and references Claude's disagreement with Anthropic's redaction in the report.

The author describes the common enterprise problem where PII redaction before sending data to LLMs breaks the output, and they are building a solution that rehydrates responses without exposing raw data.

This post presents an optimized method for full document redaction using Qwen 3.6 27B with a Pi agent harness, achieving acceptable results through higher quantization, improved prompting, and a Gradio UI.

Introduces ScreenLeak, a benchmark for measuring PII redaction in computer-use AI data, and presents two local models (v45_phase3 for text and rfdetr_v8 for images) achieving near-frontier performance at low latency.

Whiteout is a macOS tool that automatically redacts sensitive information from screenshots to protect privacy.

BlankOut is a tool that redacts sensitive content in documents on-device before sharing with AI services.